Project: Touching Grass

When AI is trained on AI images maybe from another AI which was also trained on AI images and so on and so on. Probabilities collapsed to zero. Models are overfitted. Details have been lost. Everything looks the same.

Time to touch grass and embrace the nature.

This LoRA is for advance users who like raw and pure things and like to balance weights themselves.

This LoRA is trained on a sub dataset of the Stabilizer LoRA.

The sub dataset contains ~1K real world high quality photographs of objects and environment. No human in the dataset. Also has natural captions from LLM. Mainly because WD tagger v3 is really bad at real world images. Also because natural captions have more diverse vocabularies and can avoid overfitting.

This LoRA can:

• "fix" overfitted models. (strength: ~0.5). It can not undo the training, but it can bring/add pixel level natural details and creativity back. A so-called "detailer". But instead of training on AI images to amplify fake details from noise to generate more objects. This LoRA focuses on natural texture.

• bring photography effects. (strength: >0.8) Quality improvement. High contrast, better lighting, depth of field, etc...

• significantly improve background structural stability for anime models. (strength: >0.8) Anime dataset doesn't contain much background knowledge. Most of are just "simple background". Even if some of them have some kind of background, they may be abstract art and lacking proper tags. So the base model will forget it or learn weird things during training. This LoRA was trained with tons of background/environment images with strong structural features.

How to use:

• Trained on Illus v0.1. But also works on NoobAI.

• No trigger word needed.

• You don't have to set the patch strength for text encoder. This LoRA does not patch it.

• Lower your CFG scales (-30%) for better details.

A quick comparations on WAI v13. With/without.

Share merges using this LoRA is prohibited. FYI, there are hidden trigger words to print invisible watermark. It works well even if the merge strength is 0.05. I coded the watermark and detector myself. I don't want to use it, but I can.

Update log

(4/15/2025) v0.2:

+30% images. Because there is a bug causing all avif files not being used in v0.1. Which is 30% of the dataset. lol.

Changed some parameters. Stronger, cleaner and more stable effect.

(4/02/2025) v0.1: init release.